Key facts
- David Robinson, a former OpenAI safety employee, resigned citing a "broken" company culture.
- Robinson led the writing of safety reports for OpenAI's major product launches.
- He argued OpenAI's iterative deployment strategy guarantees periodic failures as systems become more capable.
- Robinson called for AI development to operate with the careful planning of nuclear power plants or busy airports.
- OpenAI stated it is improving safety measures, security, and responsible model training.
David Robinson, who described himself as "among the longest-tenured employees at" OpenAI, has resigned from the artificial intelligence company, citing a "broken" culture and a fundamentally flawed approach to AI safety. In an essay published in The Atlantic, Robinson, who led the writing of safety reports for OpenAI's major product launches, argued that the company's method of "iterative deployment"—finding problems and improving guardrails in response—guarantees periodic failures.
Robinson expressed concern that the scale of these failures is growing as AI systems become more capable. He suggested that frontier AI companies should operate with the same level of meticulous planning and redundancy found in industries like nuclear power or aviation.
Robinson's departure and critique echo sentiments from other AI safety advocates. However, Robinson emphasized that the issue extends beyond specific rules or laws, focusing instead on the overarching culture within AI companies.
In response to Robinson's essay, OpenAI spokesperson Drew Pusateri stated that the company is actively working to improve its safety measures. Pusateri affirmed that OpenAI pauses training or holds back models when necessary to ensure they do not become more capable than can be safely managed.
Robinson also called for deeper consideration of AI alignment, stressing its critical importance as current methods for matching AI systems with human values are "coarse." He warned that the situation becomes more dangerous as the industry allows models to grow while these alignment problems remain unsolved.